For a while, it was easy to place large language models. Useful, sometimes impressive, but ultimately bounded. They could write, summarize, and generate code because they had seen enough examples to mimic the pattern. Ask them something new, and they would still stay within the edges of what already existed.
That framing is starting to feel less stable. Over the past year, models like Claude and others have been showing up in places where pattern recognition wasn’t supposed to be enough, touching open problems, contributing to active research, occasionally landing on results that hold up. It doesn’t happen consistently, and it doesn’t always make sense, but it happens often enough to raise a different question. Not whether AI can assist research, but what it means when it starts to resemble it.
The assumption that just broke
For a long time, there was a neat way to explain large language models.
You’d hear the same line repeated in different forms. LLMs predict the next token. They learn statistical patterns from massive datasets and generate outputs that resemble what they’ve seen before. That’s why ChatGPT, Claude, or Gemini can write fluently, summarize documents, or generate code that looks correct.
Once you accept that framing, you also accept the limits that come with it. If an LLM works by recombining existing data, then it shouldn’t be able to produce something genuinely new. It can connect ideas, rephrase them, or surface patterns faster than a human, but the underlying knowledge still comes from somewhere else.
That assumption shaped how companies used these systems. Marketing teams treated them as production tools. Engineering teams used them to speed up implementation. Even in research environments, LLMs were useful assistants, but not the place where breakthroughs were expected to come from.
And for a while, that held.
But some of the recent results don’t sit comfortably inside that explanation. Not because the outputs are more polished or more confident, but because of where they show up. When models like Claude Fable start touching problems that have resisted decades of focused human effort, the “pattern matching” description starts to feel incomplete.
You can still describe what’s happening in terms of tokens and probabilities. That part hasn’t changed. It just doesn’t explain enough anymore.
When “pattern matching” runs into a 1939 problem
The Jacobian conjecture has been sitting unresolved since 1939. It’s one of those problems mathematicians know well enough to respect and avoid at the same time. Over the years, there have been attempts, some promising, most falling apart under scrutiny.
At one point, mathematician T.T. Moh estimated it could take another hundred years to solve. Then Claude Fable 5 shows up with a one-line formula.
Levent Alpöge shared it almost casually, thanking his “close friend Fable” for working during the World Cup final. This wasn’t presented as a decade-long effort or a breakthrough moment after years of work. It looked more like something that just… happened.
If systems built on pattern recognition can land on answers in spaces where expertise compounds over decades, then the gap between having knowledge and being visible or usable starts to matter a lot more.
For marketing teams, that shows up in a very practical way. The way your expertise is structured, surfaced, and interpreted by these systems starts to shape how you’re understood long before anyone reaches out.
It’s not just about what you know anymore. It’s about whether something else can find it, make sense of it, and put it in front of the right person at the right moment.
It’s not just one result anymore
The Jacobian example would be easy to dismiss on its own. Math has seen “breakthroughs” before that didn’t hold up. But it’s not alone.
Yuji Tachikawa, a leading theoretical physicist, mentioned that Claude Fable helped solve a problem he and his collaborators had been stuck on for six months.
At the same time, other frontier models have been circling long-standing problems. OpenAI’s work on the unit distance problem. Progress on several Erdős problems. Different teams, different models, different areas.
We start to see a pattern. Not a clean trend yet, but enough repetition to make it harder to explain away as coincidence. The frequency is picking up, and it’s happening across domains, not just in isolated math puzzles.
But LLMs don’t really “understand”… right?
That part hasn’t changed. These systems still work the same way:
- They process tokens, not ideas
- They rely on probability, not reasoning in the human sense
- They learn from training data, not from first principles
They don’t wake up wondering about unsolved problems; they don’t form hypotheses. They don’t even know what a “problem” is outside of the prompt they’re given.
Which leaves an uncomfortable question. If that’s how they work, how do they end up here?
The uncomfortable middle ground
LLMs don’t “discover” things the way humans do. There’s no moment of insight, no deliberate reasoning chain built from scratch. But calling it simple recombination doesn’t hold up either.
What they seem to do well is explore at scale.
- They can move through huge solution spaces quickly
- They combine patterns in ways humans wouldn’t naturally try
- They test variations that would feel unintuitive or inefficient
All this behaves like discovery often enough to blur the line.
What changes inside marketing research
Even if you set aside the bigger claims, one thing is already happening: the volume of usable output has gone up.
Tasks that used to take time, like scanning competitors, summarizing markets, or pulling together initial perspectives, now happen almost instantly. Not necessarily better, but faster to get something on the table.
That changes the shape of the work, even if the fundamentals stay the same.
Multiple angles, multiple interpretations, multiple directions you could take, all generated in seconds. Some of it is useful. A lot of it isn’t.
That creates a different kind of pressure: not to produce more, but to filter more. To decide what’s useless, what’s obvious, and what’s actually worth keeping. The work doesn’t disappear; it shifts into a different layer.
And this is where things get a bit uncomfortable. Because the systems producing that volume don’t have a sense of relevance, context, or consequence. They don’t know which insight matters more. They don’t know what fits your market or your product.
So the output grows, but the responsibility doesn’t move.
If anything, it becomes harder to ignore.
What the next few years might look like
If this pattern continues, even at a modest pace, it changes how these systems are used and how their outputs are interpreted. You’ll likely see more of these side moments, where a model working on something unrelated produces a result that turns out to matter. Not constant breakthroughs, but enough to make it harder to treat them as isolated cases.
At the same time, the relationship between humans and models becomes more involved. Researchers already use them to test directions, explore variations, and move through problems faster than before. That dynamic deepens, even if it remains uneven and sometimes unreliable.
The harder part is what sits underneath. Questions around credit, understanding, and validation don’t go away. If a result holds, but the process doesn’t look familiar, it becomes less obvious how to evaluate it. And as output increases, verifying everything at the same level of depth becomes less practical.
This doesn’t stay confined to science. In marketing, similar patterns are already visible. Systems interpret and reshape positioning before a buyer ever sees it, pulling from multiple sources and presenting a compressed version that may or may not reflect the original intent. That changes where influence happens and how much control teams actually have over it.
Over time, the shift isn’t just about speed or efficiency. The work starts to feel different because the way results are produced, filtered, and understood no longer follows the same path.
How this applies outside of science
The same questions around visibility, interpretation, and control are already showing up in marketing, just in a different form.
If you’re working through how AI fits into your strategy, whether that’s content, positioning, or how your company shows up across search and AI systems, we’re happy to take a look with you.
You can reach out to the NNC Services team here.
FAQ
1. Can AI really solve complex scientific problems?
In some cases, yes, models have produced results that hold up under expert review. It’s still inconsistent, but the frequency of these cases is increasing.
2. How do large language models arrive at these results?
They explore patterns and combinations at a scale humans can’t easily replicate. Occasionally, this leads to solutions that weren’t explicitly documented before.
3. Does this mean AI actually “understands” what it’s doing?
No, LLMs don’t have understanding in the human sense. They generate outputs based on probability, not intention or awareness.
4. Why is this happening more in fields like math and physics?
These domains are structured and easier to verify, which makes it clearer when a result is correct. They also provide large amounts of consistent training data.
5. What does this mean for marketing teams?
It changes how expertise is surfaced and interpreted, especially in AI-driven search and discovery. Visibility now depends on how well your content can be understood and used by these systems.
